2026-06 Competitive Positioning Memo

Last refreshed: 2026-06-12. This is the public version of the strategic context referenced from ROADMAP.md. Refresh quarterly; flag anything older than two quarters as stale.

This memo exists because the ROADMAP can only be opinionated about what we will do. To explain why — and to be honest about where we fit in the 2026 TS-agent landscape — we need a separate document that quotes outside numbers, names competitors, and admits weaknesses. The audience is a prospective adopter who is comparing WasmAgent against Vercel AI SDK, LangGraph.js, OpenAI Agents JS, Mastra, or Cloudflare Agents SDK and needs to know whether to bet on us.

---

1. Honest snapshot

Technically: strong and differentiated. On distribution: nascent.

Axis	Status
Code-mode runtime	✅ Three-tier kernel matrix (in-process / WASM / remote) actually exists; competitor kernels are single-tier
Prompt-cache discipline	✅ Anthropic stable-prefix, byte-stable observer prompts, proven in ObservationalMemory
Quality runners	✅ Self-consistency / Reflect-Refine / Budget-Forcing / Parallel-Fork-Join all in core/src/enhancement/
Statistical evals	✅ McNemar exact + Wilson CI + paired bootstrap in evals-runner/src/stats (rare in JS-land)
Zero-deploy DevTools	✅ WasmAgent devtools CLI + RunsAggregator + EventLogReplay (no SaaS, no account)
npm distribution	⚠️ @wasmagent/core first published 2026-06-12 — still a single version on npm
Bus factor	⚠️ Single maintainer; no co-publishers yet
Public benchmarks	⚠️ Internal LongMemEval-style runs published; the official 500-question set has not yet been run

The technical foundation is real. The market signal is not yet visible. This memo is about how we close that gap without spending the next year reinventing what Vercel/Mastra already ship.

2. Where the market actually sits in 2026-06

Three observations from the field, all later than our own history:

1. "Code-mode" has become the consensus pattern. Cloudflare's 2026-02 Code Mode: give agents an entire API in 1,000 tokens blog, the 2026-04 Code Mode MCP server GA, and Anthropic's Code execution with MCP note all land on the same shape: expose tools as a typed code surface, run inside a sandbox, collapse the per-tool round-trip cost. This validates S1 — the axis we picked for the runtime.
2. The framework race already has a winner. Vercel AI SDK, by downloads, is roughly four orders of magnitude ahead of every other TS-native framework. LangGraph.js is the enterprise default. Mastra closed a 2026-04 Series A on the back of LongMemEval and their Studio product. There is no remaining oxygen for a "new framework" pitch — there is oxygen for a runtime that the leaders embed.
3. Selection criteria are tilting toward governance signals. Pharos, MAG, Alice Labs, and the Forrester MCP brief all weight release cadence, security response, bus factor, and integration breadth equally with raw features. A repo with one npm version and one maintainer fails most enterprise checklists before the technical merits are scored.

3. The three strategic lines

The ROADMAP's S1–S4 stay; this memo refines them into three lines the next two quarters of work must serve:

• L1 — Become the embedded runtime. Stop pitching the framework. Pitch the kernel + the manifest + the evaluator. Ship adapters into Vercel AI SDK, Mastra, Claude Agent SDK, and OpenAI Agents JS, and request inclusion in their official provider lists.
• L2 — Trade self-built numbers for public-leaderboard numbers. Mastra leveraged a single LongMemEval headline into the entire press cycle. Our reply is a Pareto report (accuracy × cost × latency × token-efficiency × estimated J/correct) that no single-axis headline can dominate. Pareto is harder to spin and harder to ignore.
• L3 — Make trust legible. A co-maintainer, a CHANGELOG with a cadence, a 1.0 freeze face, and a public security response history do more for adoption today than any new feature would. They are cheap. We have not done them yet.

4. What this memo refuses to do

• No new framework features whose only justification is "Vercel / Mastra has it." Feature parity is a losing race when distribution is four orders of magnitude apart. Either the feature is useful to us as a runtime, or it does not ship.
• No private benchmark headlines. If a number cannot be reproduced by examples/benchmarks/*.mjs against a public dataset, it does not appear in README.
• No promises we cannot resource. Anything requiring sustained cloud spend (D2 LongMemEval-500 full run, paid security audit) is flagged 🖥️ in the ROADMAP so adopters know it depends on funding.

5. How to challenge this memo

The single fastest way to falsify this strategy is to find that adopters do not care about a neutral, multi-language, multi-isolation code-mode runtime — that they are content with whatever sandbox Cloudflare or Anthropic ships first, even with the platform lock-in. If by 2026-Q4 the upstream-adapter packages (aisdk, mastra-sandbox, plus the ones added under D1) show zero organic downloads from the upstream ecosystems, this memo is wrong and the runtime pitch should be retired in favor of either a pure DevTools play or sunset.

Watch the evals reports directory and the upstream adapter download graphs; that is the signal.